Stanford Vision and Learning Lab

Mobile manipulation policies can break from base pose errors of just a few cm, which are common after navigation. Can we get pose generalization without additional demos? Introducing MobileVISTA: collect demos at one base pose, get a policy that works from many 🧵👇

Website: sdwistreich.github.io/mobilevista
Paper: arxiv.org/abs/2610.07511

Stanford Vision and Learning Lab members and alumni: we are organizing a SVL+Alumni meet up at CVPR’24! Please check your emails for more details.

Stanford Vision and Learning Lab is presenting 7 papers at #CORL2023, including 3 oral presentations, and 3 award nominations, see below:

RL in multitask domains with shared underlying latent information can be made more effective through learning action space manifolds.
Check out this effort on Learning latent Action Spaces at #ICRA2021

We are hosting one of the 3 challenges of embodied-ai.org at CVPR20. Train your navigating agent in our simulator Gibson (svl.stanford.edu/gibson2) and we will test it in the real world! The best solutions will showcase live during CVPR. More info: svl.stanford.edu/gibson2/challe…

Introducing the RoboTurk Real Robot Dataset - one of the largest, richest, and most diverse robot manipulation datasets ever collected using human creativity and dexterity!
111 hours
54 non-expert demonstrators
2144 demonstrations
Download: roboturk.stanford.edu/realrobotdatas…
[1/2]

Come check out the poster on recent work in imitation learning at #NeurIPS

Our focus on robot learning from single example of a task through a video has resulted in a line of work that combines symbolic systems with neural networks

This is continuing an important line of work in policy learning with large Datasets.

More importantly this is the question to answer if we need to create a analog of "Imagenet" for Robotics.

We need to both collect large Datasets and have Algorithms to leverage this data!

Continued efforts in larger scale crowdsourcing robotics for dataset creation in setups where engineered solutions are hard, simulation is tricky and pure compute has low success. The diversity of human cognitive reasoning and dexterity provides so many ways do do the same task!

A new work on structuring diverse semantics in 3D space that yielded the 3D Scene Graph! It’s showcased on the Gibson database by annotating the models with diverse semantics using a semi-automated method.

Learning from hints (not demonstrations): A new paper on an important direction of RL for control where expert intuition can be used to guide learning without the need to provide optimal or even complete solutions.

[1/2] Our lab has 3 papers accepted to NeurIPS 2019:
1. HYPE: Human Eye Perceptual Evaluation of Generative Models. Zhou and Gordon et al. (Oral)
2. SOCIAL-BIGAT: Multimodal Trajectory Forecasting using Bicycle-GAN and Graph Attention Networks. Kosaraju et al.

A new effort from SVL and the JackRabbot team! New dataset and benchmark for robot perception in human environments. The winners of the first challenge on pedestrian detection and tracking will be presented at our workshop at #ICCV19! jrdb.stanford.edu/workshops/jrdb…

We are happy to announce our ICCV19 Workshop on Visual Perception for Navigation in Human Environments: The JackRabbot Social Robotics Dataset and Benchmark. Submission deadline August 20. jrdb.stanford.edu/workshops/jrdb… For more info, contact @SHamidRezatofig and Roberto Martin-Martin

A thorough evaluation of possible action spaces and their efficiency with RL for manipulation tasks. The answer is best of both worlds - Op. Space control in end effector space with learnable gains fares better than end to end image-to-torque

Stanford Vision and Learning Lab: Performing Research at the Forefront of Computer Vision, Machine Learning, and Robotics - HostingAdvice.com
⁦@drfeifei⁩ ⁦@silviocinguetta⁩ ⁦@jcniebles⁩ hostingadvice.com/blog/a-look-at…

New study that analyzes different action spaces for RL in robot manipulation in the quest to find the best one. Guys what? The best method is a combination of operational space control (1986) with learned adaptive gain tuning. To appear at #IROS2019

Lab PI @drfeifei wins the PAMI Longuest-Higgins Prize for the effort on ImageNet with @lijiali_vision @jiadeng @RichardSocher and team. Congrats everyone

Exciting new line of work from the group on 3D vision at #CVPR2019 by @lynetcha @SHamidRezatofig @silviocinguetta